Data analytics for supporting clinical decision on patients with implantable cardioverter defibrillator

Abstract

Nowadays, the massive storage of cardiac arrhythmic episodes from Implantable Cardioverter Defribrillators (ICDs) is opening up a new range of opportunities for electrophysiological knowledge extraction. Large and high quality databases are increasingly encouraging the development of new Data Analytics (DA) tools supporting cardiologists on their clinical decisions. Within this new context, this Thesis aims to provide a computational solution for two current challenges in cardiology: (1) the automatic classification of cardiac arrhythmic episodes recorded by ICDs; and (2) the determination of a safety threshold on R-wave amplitudes during a normal Baseline Rhythm (BR) for ensuring a low risk of undersensing fatal arrhythmic episodes in ICDs. On the one hand, current ICDs are highly reliable detecting fatal arrhythmic episodes. However, the accurate automatic classification into specific classes remains a field for improvement. As a result, cardiologists still need to manually analyze each episode and to check whether the ICD detection and treatment were adequate. Therefore, a novel DA-based methodology for the automatic classification of ICD arrhythmic episodes is proposed in this Thesis. The methodology is defined to be potentially used in real-world ICD scenarios, since: (1) it requires minimal signal preprocessing due to memory and battery constraints; and (2) it deals with episodes of different duration in the presence of non-recording intervals. Likewise, the proposed methodology emulates the know-how of expert cardiologists, for which it simultaneously considers heart activation events and signal waveforms. Results on a set of 6,233 actual ICD detected episodes from 599 patients showed test accuracy rate (and kappa coefficient) close to 78% (0.6) and 90% (0.8) in both 8 and 3-class imbalanced schemes, respectively. On the other hand, R-wave amplitudes during normal BR are the current indicators used to subjectively characterize the risk of undersensing Ventricular Fibrillation (VF) episodes in ICDs. However, a minimum value (or safety threshold) has not been established yet. Clinical guidelines recommend R-wave amplitudes of at least 7 mV at the ICD implantation. When the amplitude is lower, it is usual the induction of defibrillation tests in the clinical practice to ensure that undersensing does not occur. The drawback of this is those inductions increase the patient complications, and the efficacy and safety of ICD therapies are not improved by using these data as secondary information. In order to tackle this challenge, a DA-based procedure for estimating a safety threshold on BR R-wave amplitudes is proposed in this Thesis. To define this procedure: (1) the behavior and undersensing rate of R-wave amplitudes during VF episodes are defined; and (2) the R-wave amplitude relationship between BR and VF episodes is determined. Results on a set of 229 actual VF episodes from 83 patients showed that R-wave amplitudes lower than 2.47 mV can lead to potentially risk situations of non or late detection of VF episodes. This Thesis contributes to scientific literature by offering new insights for the development of new DA-based tools to support cardiologists during the follow-up of patients with an ICD. Results for both raised challenges convincingly demonstrate that the new generation of large and high quality clinical databases plays a major role in future trends in cardiology.